Model Card for Gemma-4-31B-storymaxxed3
This model is a fine-tuned version of llmfan46/gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic. It has been trained using TRL. Optimized specifically for creative writing and narrative prose.
The same dataset was used as that for storymaxxed 1 and 2, but with a different base model this time: llmfan46/gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic along with some tweaks to the training setup.
The first two storymaxxed models needed more writerly prose in my opinion and Ortenzya has excellent word selection and overall just a better writer than stock Gemma-4. I think it turned out rather well, with good dialogue generation and excellent scene and detail tracking.
This model should exhibit the same prowess with producing quality stories/narratives with improved prose and better dialogue compared to Storymaxxed 1 and 2.
Training procedure
This model was trained with TRL using DPO on a high quality dataset of narrative preference pairs.
Introduction to training method used: Direct Preference Optimization: Your Language Model is Secretly a Reward Model.
Recommended Sampler Settings
For optimal inference, use the standard generation parameters recommended by Google for Gemma-4 models:
- Temperature - 1.0
- Top P - 0.95
- Top K - 64
Vision mmproj
The mmproj file for vision can be found here: https://huggingface.co/MRockatansky/Gemma-4-31B-Storymaxxed3-GGUF